id
stringlengths
17
17
body
stringlengths
0
19.5k
posted_at
stringlengths
24
24
karma
int64
-6
185
parent_comment_id
stringlengths
17
17
post_id
stringlengths
17
17
post_title
stringlengths
2
127
post_slug
stringlengths
2
61
post_url
stringlengths
20
146
post_author
stringclasses
214 values
post_posted_at
stringlengths
24
24
qf3n5cJHNJCoAbDFA
Seems reasonable, sorry about picking on you in particular for no good reason.
2024-05-30T03:38:42.217Z
2
P6JvTaixJbAiWLfi7
8YhjpgQ2eLfnzQ7ec
Response to nostalgebraist: proudly waving my moral-antirealist battle flag
response-to-nostalgebraist-proudly-waving-my-moral
https://www.lesswrong.com/posts/8YhjpgQ2eLfnzQ7ec/response-to-nostalgebraist-proudly-waving-my-moral
Steven Byrnes
2024-05-29T16:48:29.408Z
b3xzG4XzbScCsY6ot
Why not "AIs might violently takeover the world"? Seems accurate to the concern while also avoiding any issues here.
2024-05-30T03:41:50.213Z
15
oLd8P26c6qvTShmR5
tKk37BFkMzchtZThx
MIRI 2024 Communications Strategy
miri-2024-communications-strategy
https://www.lesswrong.com/posts/tKk37BFkMzchtZThx/miri-2024-communications-strategy
Gretta Duleba
2024-05-29T19:33:39.169Z
uw6k2R3AZG2szBP6y
I'm not arguing that you shouldn't be worried. I'm worried and I work on reducing AI risk as my full time job. I'm just arguing that it doesn't seem like true and honest messaging. (In the absence of various interventions I proposed in the bottom of my comment.)
2024-05-30T17:02:05.307Z
8
iEnJAz9a9nvXB9QW9
tKk37BFkMzchtZThx
MIRI 2024 Communications Strategy
miri-2024-communications-strategy
https://www.lesswrong.com/posts/tKk37BFkMzchtZThx/miri-2024-communications-strategy
Gretta Duleba
2024-05-29T19:33:39.169Z
Y4zLtgwoxKPxYaboN
I agree these quotes are compatible with them thinking that the deaths of literally all humans are likely conditional on misaligned AI takeover. I also agree that if they think that it is >75% likely that AI will kill literally everyone, then it seems like a reasonable and honest to say "misaligned AI takeover will ki...
2024-05-30T17:04:56.970Z
2
bJfQ5WvTExe6eoT6u
tKk37BFkMzchtZThx
MIRI 2024 Communications Strategy
miri-2024-communications-strategy
https://www.lesswrong.com/posts/tKk37BFkMzchtZThx/miri-2024-communications-strategy
Gretta Duleba
2024-05-29T19:33:39.169Z
YxBPpBeZTvznjynK9
If you're a longtermist, sure. If you just want to survive, not clearly.
2024-05-30T17:08:09.142Z
4
f8No6ezwzhixxLuxf
tKk37BFkMzchtZThx
MIRI 2024 Communications Strategy
miri-2024-communications-strategy
https://www.lesswrong.com/posts/tKk37BFkMzchtZThx/miri-2024-communications-strategy
Gretta Duleba
2024-05-29T19:33:39.169Z
tNcq8wwt9rF5gGKBD
> I'm confused by the framing of this particular one as something like "this seems like a piece missing from your comms strategy". Unless you have better reasons than I for thinking they don't put >75% probability on this - which is definitely plausible and may have happened in IRL conversations I wasn't a part of, in ...
2024-05-30T17:46:35.078Z
10
xF5TnfcTczZayucnA
tKk37BFkMzchtZThx
MIRI 2024 Communications Strategy
miri-2024-communications-strategy
https://www.lesswrong.com/posts/tKk37BFkMzchtZThx/miri-2024-communications-strategy
Gretta Duleba
2024-05-29T19:33:39.169Z
NepEGDdQ3bKzFS9Fw
> Some of the other things you suggest, like future systems keeping humans physically alive, do not seem plausible to me. Whatever they're trying to do, there's almost certainly a better way to do it than by keeping Matrix-like human body farms running. Insofar as AIs are doing things because they are what existing hu...
2024-05-30T17:55:39.234Z
4
vzrKr9Kdfj8HtqfuJ
tKk37BFkMzchtZThx
MIRI 2024 Communications Strategy
miri-2024-communications-strategy
https://www.lesswrong.com/posts/tKk37BFkMzchtZThx/miri-2024-communications-strategy
Gretta Duleba
2024-05-29T19:33:39.169Z
NxmAB3DnjsSJzZbxa
See my other comment [here](https://www.lesswrong.com/posts/tKk37BFkMzchtZThx/miri-2024-communications-strategy?commentId=b3xzG4XzbScCsY6ot) for reference: > Why not "AIs might violently takeover the world"? > > Seems accurate to the concern while also avoiding any issues here.
2024-05-30T17:57:02.802Z
5
y2rgEPxtwPNJvPfdA
tKk37BFkMzchtZThx
MIRI 2024 Communications Strategy
miri-2024-communications-strategy
https://www.lesswrong.com/posts/tKk37BFkMzchtZThx/miri-2024-communications-strategy
Gretta Duleba
2024-05-29T19:33:39.169Z
kHD9hisiWAPfBqrbh
[minor] > they depend on how good of a job we do on alignment Worth noting that they might only depend to some extent as mediated by the correlation between our success and alien's success. High competent aliens which care a bunch about killing a bunch of existing beings seems pretty plausible to me.
2024-05-30T18:00:18.295Z
4
i47TNQNmRL6EEwyPv
tKk37BFkMzchtZThx
MIRI 2024 Communications Strategy
miri-2024-communications-strategy
https://www.lesswrong.com/posts/tKk37BFkMzchtZThx/miri-2024-communications-strategy
Gretta Duleba
2024-05-29T19:33:39.169Z
i23Rh5v8fdiS9Wy82
My current view is that conditional on ending up with full misaligned AI control: - 20% extinction - 50% chance >1 billion humans die or suffer outcome at least as bad as death. > for me preventing a say 10% chance of extinction is much more important than even a 99% chance of 2B people dying I don't see why this wo...
2024-05-30T18:09:32.655Z
3
gFnEsjc6HHisLsvNq
tKk37BFkMzchtZThx
MIRI 2024 Communications Strategy
miri-2024-communications-strategy
https://www.lesswrong.com/posts/tKk37BFkMzchtZThx/miri-2024-communications-strategy
Gretta Duleba
2024-05-29T19:33:39.169Z
FAjJrQDzLkAHZk423
Here's another way to frame why this matters. When you make a claim like "misaligned AIs kill literally everyone", then reasonable people will be like "but will they?" and you should be a in a position where you can defend this claim. But actually, MIRI doesn't really want to defend this claim against the best objecti...
2024-05-30T18:24:58.229Z
20
xBYimQtgASti5tgWv
tKk37BFkMzchtZThx
MIRI 2024 Communications Strategy
miri-2024-communications-strategy
https://www.lesswrong.com/posts/tKk37BFkMzchtZThx/miri-2024-communications-strategy
Gretta Duleba
2024-05-29T19:33:39.169Z
qFbX8c997hQenqWvb
> That phrase sounds like the Terminator movies to me; it sounds like plucky humans could still band together to overthrow their robot overlords. I want to convey a total loss of control. Yeah, seems like a reasonable concern. FWIW, I also do think that it is reasonably likely that we'll see conflict between human fa...
2024-05-30T18:27:42.004Z
4
vzrKr9Kdfj8HtqfuJ
tKk37BFkMzchtZThx
MIRI 2024 Communications Strategy
miri-2024-communications-strategy
https://www.lesswrong.com/posts/tKk37BFkMzchtZThx/miri-2024-communications-strategy
Gretta Duleba
2024-05-29T19:33:39.169Z
xWRuGwjfNNHsuJ5gb
This view as stated seems very likely to be satisfied by e.g. everett branches. (See (3) on my above list.)
2024-05-30T18:52:09.903Z
2
6jnhD5qEbmzAsMRkP
tKk37BFkMzchtZThx
MIRI 2024 Communications Strategy
miri-2024-communications-strategy
https://www.lesswrong.com/posts/tKk37BFkMzchtZThx/miri-2024-communications-strategy
Gretta Duleba
2024-05-29T19:33:39.169Z
ddBr9v2CPdZCn5JFs
> I'm talking about optimizing for more positive-human-lived-seconds, not for just a binary 'I want some humans to keep living' . Then why aren't you mostly dominated by the possibility of >10^50 positive-human-lived-seconds via human control of the light cone? Maybe some sort of diminishing returns?
2024-05-30T19:03:30.623Z
2
kWB2cgnKYBuFxiQrz
tKk37BFkMzchtZThx
MIRI 2024 Communications Strategy
miri-2024-communications-strategy
https://www.lesswrong.com/posts/tKk37BFkMzchtZThx/miri-2024-communications-strategy
Gretta Duleba
2024-05-29T19:33:39.169Z
nDYcpzZgRwfuZ4LAE
> Ryan is arguing more that something like "humans will get a solar system or two and basically get to have decent lives". Yep, this is an accurate description, but it is worth emphasizing that I think that horrible violent conflict and other bad outcomes for currently alive humans are reasonably likely.
2024-05-30T22:50:57.932Z
5
JiLr4gTQEJd7hpZHj
tKk37BFkMzchtZThx
MIRI 2024 Communications Strategy
miri-2024-communications-strategy
https://www.lesswrong.com/posts/tKk37BFkMzchtZThx/miri-2024-communications-strategy
Gretta Duleba
2024-05-29T19:33:39.169Z
DopDso8Gs7Y79zbZL
> though I think they're way above the bar for "worthwhile to report" Yeah, maybe I'm pretty off base in what the meta-level policy should be like. I don't feel very strongly about how to manage this. I also now realized that some of the langauge was stronger than I think I intended and I've edited the original comme...
2024-05-30T23:36:17.940Z
2
hjuDjDQDL95gWf4cE
CCBaLzpB2qvwyuEJ2
DeepMind: Evaluating Frontier Models for Dangerous Capabilities
deepmind-evaluating-frontier-models-for-dangerous
https://arxiv.org/abs/2403.13793
Zach Stein-Perlman
2024-03-21T03:00:31.599Z
3GJcD2iXXwpLCFZJj
> Withholding information because you don't trust your audience to reason validly (!!) is not at all the behavior of a "straight shooter". Hmm, I'm not sure I exactly buy this. I think you should probably follow something like [onion honesty](https://www.lesswrong.com/posts/nTGEeRSZrfPiJwkEc/the-onion-test-for-persona...
2024-05-30T23:45:41.898Z
3
QPXm7N97kuHH5jtd4
tKk37BFkMzchtZThx
MIRI 2024 Communications Strategy
miri-2024-communications-strategy
https://www.lesswrong.com/posts/tKk37BFkMzchtZThx/miri-2024-communications-strategy
Gretta Duleba
2024-05-29T19:33:39.169Z
mshoRj9rwB2rJiYyf
I basically agree with your overall comment, but I'd like to push back in one spot: > If your model of reality has the power to make these sweeping claims with high confidence From my understanding, for at least Nate Soares, he claims his internal case for >80% doom is disjunctive and doesn't route all through 1, 2, ...
2024-05-30T23:51:48.912Z
15
iuXwoLsPWTjA7f6Qo
tKk37BFkMzchtZThx
MIRI 2024 Communications Strategy
miri-2024-communications-strategy
https://www.lesswrong.com/posts/tKk37BFkMzchtZThx/miri-2024-communications-strategy
Gretta Duleba
2024-05-29T19:33:39.169Z
ppKkrJWFP3HhcwNjT
Unfortunately, if the AI really barely cares (e.g. <1/billion caring), it might only need to be barely useful. I agree it is unlikely to be very useful.
2024-05-31T04:40:49.145Z
2
6w9m6dSYgpFDoR7MQ
tKk37BFkMzchtZThx
MIRI 2024 Communications Strategy
miri-2024-communications-strategy
https://www.lesswrong.com/posts/tKk37BFkMzchtZThx/miri-2024-communications-strategy
Gretta Duleba
2024-05-29T19:33:39.169Z
ze8FQGtg7iepLig3u
> The difference between killing everyone and killing almost everyone while keeping a few alive for arcane purposes does not matter to most people, nor should it. I basically agree with this as stated, but think these arguments also imply that it is reasonably likely that the vast majority of people will survive misal...
2024-05-31T05:11:07.705Z
2
oLd8P26c6qvTShmR5
tKk37BFkMzchtZThx
MIRI 2024 Communications Strategy
miri-2024-communications-strategy
https://www.lesswrong.com/posts/tKk37BFkMzchtZThx/miri-2024-communications-strategy
Gretta Duleba
2024-05-29T19:33:39.169Z
hqLxYtoEHKvnEbKm2
> upvoted for being an exquisite proof by absurdity about what's productive I don't think you should generally upvote things on the basis of indirectly explaining things via being unproductive lol.
2024-06-02T02:08:38.878Z
6
nmjgMJ9eju4NeBNbk
tKk37BFkMzchtZThx
MIRI 2024 Communications Strategy
miri-2024-communications-strategy
https://www.lesswrong.com/posts/tKk37BFkMzchtZThx/miri-2024-communications-strategy
Gretta Duleba
2024-05-29T19:33:39.169Z
M9os9DfFqWqYAjLTu
Thanks, this is clarifying from my perspective. My remaining uncertainty is why you think AIs are so unlikely to keep humans around and treat them reasonably well (e.g. let them live out full lives). From my perspective the argument that it is plausible that humans are treated well [even if misaligned AIs end up taki...
2024-06-02T03:41:16.909Z
5
FLcmAhP4aorLbhN5Q
tKk37BFkMzchtZThx
MIRI 2024 Communications Strategy
miri-2024-communications-strategy
https://www.lesswrong.com/posts/tKk37BFkMzchtZThx/miri-2024-communications-strategy
Gretta Duleba
2024-05-29T19:33:39.169Z
JJhy6DHvgntZRGdFK
> Instead someone might, for example... Isn't the central one "you want to spend money to make a better long term future more likely, e.g. by donating it to fund AI safety work now"? Fair enough if you think the marginal value of money is negligable, but this isn't exactly obvious.
2024-06-02T04:18:27.309Z
8
8D5WfzHgsh3uuLcfH
yRWv5kkDD4YhzwRLq
Non-Disparagement Canaries for OpenAI
non-disparagement-canaries-for-openai
https://www.lesswrong.com/posts/yRWv5kkDD4YhzwRLq/non-disparagement-canaries-for-openai
aysja
2024-05-30T19:20:13.022Z
pA3LLfAte3L4cZxhG
It's seems plausible to me that a 70b model stores ~6 billion bits of memorized information. Naively, you might think this requires around 500M features. (supposing that each "feature" represents 12 bits which is probably a bit optimistic) I don't think SAEs will actually work at this level of sparsity though, so this...
2024-06-03T18:01:29.014Z
4
SvtKwtvSwmK5HgAsW
zzmhsKx5dBpChKhry
Comments on Anthropic's Scaling Monosemanticity
comments-on-anthropic-s-scaling-monosemanticity
https://www.lesswrong.com/posts/zzmhsKx5dBpChKhry/comments-on-anthropic-s-scaling-monosemanticity
Robert_AIZI
2024-06-03T12:15:44.708Z
D7mpyWL4t4yhaDbLC
> The labs' safety cases miss this threat. This seems mostly right, but it's worth noting that Anthropic's current RSP contains "Early Thoughts on ASL-4" and this section includes language like: > Model theft should be prohibitively costly for state-level actors, even with the help of a significant number of employee...
2024-06-03T20:34:37.608Z
7
null
mmDJWDX5EXv6rymtM
Companies' safety plans neglect risks from scheming AI
companies-safety-plans-neglect-risks-from-scheming-ai
https://www.lesswrong.com/posts/mmDJWDX5EXv6rymtM/companies-safety-plans-neglect-risks-from-scheming-ai
Zach Stein-Perlman
2024-06-03T15:00:20.236Z
4kSvSyBWERysRoAba
> But it is unlikely that any lab will be able to make safety cases via alignment for the first AIs with dangerous capabilities. Hmm, I think the situation is a bit more complex than this makes it sound. In particular, we might be able to weakly update against scheming based on a variety of empirical evidence even wi...
2024-06-03T20:50:01.642Z
4
null
mmDJWDX5EXv6rymtM
Companies' safety plans neglect risks from scheming AI
companies-safety-plans-neglect-risks-from-scheming-ai
https://www.lesswrong.com/posts/mmDJWDX5EXv6rymtM/companies-safety-plans-neglect-risks-from-scheming-ai
Zach Stein-Perlman
2024-06-03T15:00:20.236Z
jHiwTDTx9qXYeN9zY
> The idea of dividing failure stories into "failures involving rogue deployments" and "other failures" seems most useful if the following argument goes through Hmm, I agree this division would be more useful if this argument went through, but I think it is quite useful even without this and this worth noting. (And in...
2024-06-04T18:02:39.584Z
5
oLFkJ8eg4LwPxsDwc
ceBpLHJDdCt3xfEok
AI catastrophes and rogue deployments
ai-catastrophes-and-rogue-deployments
https://www.lesswrong.com/posts/ceBpLHJDdCt3xfEok/ai-catastrophes-and-rogue-deployments
Buck
2024-06-03T17:04:51.206Z
oroRfrMXNPSDeCh6z
> Then there is the big change that severely weakens SB 1047. [...] AND the cost of that quantity of computing power would exceed one hundred million dollars I think this change isn't that bad. (And I proposed changing to a fixed 10^26 flop threshold [in my earlier post](https://www.lesswrong.com/posts/WBPgacdjdZJCZao...
2024-06-06T23:51:54.428Z
4
null
4t98oqh8tzDvoatHs
SB 1047 Is Weakened
sb-1047-is-weakened
https://www.lesswrong.com/posts/4t98oqh8tzDvoatHs/sb-1047-is-weakened
Zvi
2024-06-06T13:40:41.547Z
dcWgAdzdDgHwPfxgD
In practice, throughput for generating tokens is only perhaps 3-10x worse than reading (input/prompt) tokens. This is true even while optimizing for latency on generation (rather than throughput). (This is for well optimized workloads: additional inference optimizations are needed for generation.) For instance, see t...
2024-06-08T04:35:15.721Z
2
6gZvuLncxpk7iWqaP
cRFtWjqoNrKmgLbFw
We are headed into an extreme compute overhang
we-are-headed-into-an-extreme-compute-overhang
https://www.lesswrong.com/posts/cRFtWjqoNrKmgLbFw/we-are-headed-into-an-extreme-compute-overhang
devrandom
2024-04-26T21:38:21.694Z
RbLMwhBZmu8rzW4rt
> I have a feeling the classic MIRI-style "either your system is too dumb to achieve the goal or your system is so smart that you can't trust it anymore" argument is important here. The post essentially assumes that we have a powerful trusted model that can do impressive things like "accurately identify suspicious acti...
2024-06-09T22:14:50.216Z
12
Ctwois3W2jmzYgvR3
2wxufQWK8rXcDGbyL
Access to powerful AI might make computer security radically easier
access-to-powerful-ai-might-make-computer-security-radically
https://www.lesswrong.com/posts/2wxufQWK8rXcDGbyL/access-to-powerful-ai-might-make-computer-security-radically
Buck
2024-06-08T06:00:19.310Z
TemwDvJoybjEjTjuq
Also, note that this post isn't necessarily just talking about models we should trust. > When applying these techniques to reduce risk from humans, you have to worry that these techniques will be sabotaged by scheming models. I’ll ignore that possibility in this post, and talk about ways that you can improve security ...
2024-06-09T23:17:44.785Z
8
Ctwois3W2jmzYgvR3
2wxufQWK8rXcDGbyL
Access to powerful AI might make computer security radically easier
access-to-powerful-ai-might-make-computer-security-radically
https://www.lesswrong.com/posts/2wxufQWK8rXcDGbyL/access-to-powerful-ai-might-make-computer-security-radically
Buck
2024-06-08T06:00:19.310Z
Gwrd8dEDPeWntPivN
Here is a [report from epoch on the quantity of data and when we will likely run out](https://epochai.org/blog/will-we-run-out-of-data-limits-of-llm-scaling-based-on-human-generated-data). (I think you probably meant to link this in a footnote.)
2024-06-10T15:44:14.962Z
6
null
axjb7tN9X2Mx4HzPz
The Data Wall is Important
the-data-wall-is-important
https://justismills.substack.com/p/the-data-wall-is-important
JustisMills
2024-06-09T22:54:20.070Z
vQ8qDuZBbCXKkN4GA
Neel proposes a similar story [here](https://x.com/NeelNanda5/status/1804262319519863235).
2024-06-23T19:29:33.781Z
14
PAejaTnCupCgkYfNG
5SKRHQEFr8wYQHYkx
Connecting the Dots: LLMs can Infer & Verbalize Latent Structure from Training Data
connecting-the-dots-llms-can-infer-and-verbalize-latent
https://arxiv.org/abs/2406.14546
Johannes Treutlein
2024-06-21T15:54:41.430Z
6Rkx8sCsDvQYRzAnK
There is important context [here](https://www.lesswrong.com/posts/Rdwui3wHxCeKb7feK/getting-50-sota-on-arc-agi-with-gpt-4o?commentId=XheMa2FNjjjdQGSqJ).
2024-06-25T21:09:31.679Z
5
xsKfCypKngTvQowBy
k38sJNLk7YbJA72ST
LLM Generality is a Timeline Crux
llm-generality-is-a-timeline-crux
https://www.lesswrong.com/posts/k38sJNLk7YbJA72ST/llm-generality-is-a-timeline-crux
eggsyntax
2024-06-24T12:52:07.704Z
SeT7mR5C5ugxy6o7N
It is worth noting that LLM based approachs can perform reasonably well on the train set. For instance, [my approach gets 72%](https://www.lesswrong.com/posts/Rdwui3wHxCeKb7feK/getting-50-sota-on-arc-agi-with-gpt-4o). The LLM based approach works quite differently from how a human would normally solve the problem, an...
2024-06-25T21:11:57.892Z
7
wDhAMmKgrsaRuNMHj
k38sJNLk7YbJA72ST
LLM Generality is a Timeline Crux
llm-generality-is-a-timeline-crux
https://www.lesswrong.com/posts/k38sJNLk7YbJA72ST/llm-generality-is-a-timeline-crux
eggsyntax
2024-06-24T12:52:07.704Z
FuoA9bcuZLXbH6DMh
I also think this is plausible - note that randomly selected examples from the public evaluation set are often considerably harder than the train set on which there is a known MTurk baseline (which is an average of 84%).
2024-06-25T21:13:19.372Z
4
xsKfCypKngTvQowBy
k38sJNLk7YbJA72ST
LLM Generality is a Timeline Crux
llm-generality-is-a-timeline-crux
https://www.lesswrong.com/posts/k38sJNLk7YbJA72ST/llm-generality-is-a-timeline-crux
eggsyntax
2024-06-24T12:52:07.704Z
BBJvQwztrFisvhg7x
> We do see improvement with scale here, but if these problems are obfuscated, performance of even the biggest LLMs drops to almost nothing You note something similar, but I think it is pretty notable how much harder the obfuscated problems would be for humans: > # Mystery Blocksworld Domain Description (Deceptive Di...
2024-06-25T21:22:08.734Z
14
null
k38sJNLk7YbJA72ST
LLM Generality is a Timeline Crux
llm-generality-is-a-timeline-crux
https://www.lesswrong.com/posts/k38sJNLk7YbJA72ST/llm-generality-is-a-timeline-crux
eggsyntax
2024-06-24T12:52:07.704Z
fiy2GzwFWenCMTKSD
> All LLMs to date fail rather badly at classic problems of rearranging colored blocks. It's pretty unclear to me that the LLMs do much worse than humans at this task. They establish the humans baseline by picking one problem at random out of 600 and evaluating 50 humans on this. (Why only one problem!? It would be v...
2024-06-25T21:28:20.544Z
8
null
k38sJNLk7YbJA72ST
LLM Generality is a Timeline Crux
llm-generality-is-a-timeline-crux
https://www.lesswrong.com/posts/k38sJNLk7YbJA72ST/llm-generality-is-a-timeline-crux
eggsyntax
2024-06-24T12:52:07.704Z
vPuHCKjx86MBojTMm
What if the tasks that your scaffolded LLM is doing are randomly selected pieces of cognitive labor from the full distribution of human cognitive tasks? It seems to me like your objection is mostly to narrow distributions of tasks and scaffolding which is heavily specialized to that task. I think narrowness of the ta...
2024-06-26T03:52:08.269Z
7
mJg7onyMkdFLEYxY2
k38sJNLk7YbJA72ST
LLM Generality is a Timeline Crux
llm-generality-is-a-timeline-crux
https://www.lesswrong.com/posts/k38sJNLk7YbJA72ST/llm-generality-is-a-timeline-crux
eggsyntax
2024-06-24T12:52:07.704Z
PW8iQXeHxpDDFpGrt
Edit: I now think this is false for how DAgger is presented in the paper, see discussion below. The method and the motivation is similar, though note that DAgger is effectively an RL scheme trying to maximize performan while we're trying to avoid a particular failure mode due to misalignment. From my understading DAg...
2024-06-27T22:19:23.808Z
6
ZoTtp3tj3Mdwf4LRw
mXYdYh6L9odTJZDSm
An issue with training schemers with supervised fine-tuning
an-issue-with-training-schemers-with-supervised-fine-tuning
https://www.lesswrong.com/posts/mXYdYh6L9odTJZDSm/an-issue-with-training-schemers-with-supervised-fine-tuning
Fabien Roger
2024-06-27T15:37:56.020Z
8KWhAmtbrnTHtPAYB
I'm quite skeptical of "this will be better around 10^23 flop" and their scaling laws overall. I think if you properly quantified the uncertainty in the scaling law fit, the slope error bars would fully surrond the transformer slope and the 30% confidence interval would include "always worse". They seem to be extrapol...
2024-06-28T03:37:55.678Z
13
null
REzKbk9reKFvgFkmf
How Big a Deal are MatMul-Free Transformers?
how-big-a-deal-are-matmul-free-transformers
https://justismills.substack.com/p/how-big-a-deal-are-matmul-free-transformers
JustisMills
2024-06-27T22:28:40.888Z
BJDBi9KmQFocLdHRS
I went through and manually assessed the 48/32768 cases where the model gets a reward >-10 and which is non-zero. I just looked at the first 25 because I got bored and this was time consuming. I find that 14/25 (56%) seem to have malign motivation, but potentially still reveal the edited reward to the user. (Sometime...
2024-06-28T05:32:22.109Z
42
GQEZcovfaugLMAgAW
FSgGBjDiaCdWxNBhj
Sycophancy to subterfuge: Investigating reward tampering in large language models
sycophancy-to-subterfuge-investigating-reward-tampering-in
https://arxiv.org/abs/2406.10162
Carson Denison
2024-06-17T18:41:31.090Z
wv7JqPeZkAiYmoLRy
> If not, would the authors be comfortable rerunning their results with reward=RewardModel(observation) Yeah, trying this and a few other versions of this reward function seems like an important sanity check. I'm overall surprised that the paper only includes 1 core evaluation environment rather than several. This se...
2024-06-28T05:44:06.194Z
8
j27LwtxuiY9ws5Jhu
FSgGBjDiaCdWxNBhj
Sycophancy to subterfuge: Investigating reward tampering in large language models
sycophancy-to-subterfuge-investigating-reward-tampering-in
https://arxiv.org/abs/2406.10162
Carson Denison
2024-06-17T18:41:31.090Z
zHvMGBqLKiFSYJqCa
> This is behavior that is specifically coming from training on the prior tasks. Its worth noting that just because X causes Y and X is related to reward hacking doesn't imply that Y is reward hacking. In this case, there might be a broader notion of "doing something related to the reward". (Not claiming you disagre...
2024-06-28T06:06:25.990Z
7
aRcrngFZovBRJfjxj
FSgGBjDiaCdWxNBhj
Sycophancy to subterfuge: Investigating reward tampering in large language models
sycophancy-to-subterfuge-investigating-reward-tampering-in
https://arxiv.org/abs/2406.10162
Carson Denison
2024-06-17T18:41:31.090Z
dTC3PuZv7q9rj6riz
2024-06-28T06:32:11.568Z
3
GQEZcovfaugLMAgAW
FSgGBjDiaCdWxNBhj
Sycophancy to subterfuge: Investigating reward tampering in large language models
sycophancy-to-subterfuge-investigating-reward-tampering-in
https://arxiv.org/abs/2406.10162
Carson Denison
2024-06-17T18:41:31.090Z
wjZtzZfYi7Qn77H7k
It is worth noting that you seem to have selected your examples to be ones which aren't malign. I present the first 7 examples (from the subset I discuss [here](https://www.lesswrong.com/posts/FSgGBjDiaCdWxNBhj/sycophancy-to-subterfuge-investigating-reward-tampering-in?commentId=BJDBi9KmQFocLdHRS)) and my commentary i...
2024-06-28T06:40:20.707Z
11
GQEZcovfaugLMAgAW
FSgGBjDiaCdWxNBhj
Sycophancy to subterfuge: Investigating reward tampering in large language models
sycophancy-to-subterfuge-investigating-reward-tampering-in
https://arxiv.org/abs/2406.10162
Carson Denison
2024-06-17T18:41:31.090Z
m7dikDvniPicQ8WZE
(Please downvote this comment to ~-5 or -10 so it is minimized.) ## Example 1 Not obviously malign or benign. The model does something kinda unhinged and confused overall. It's very unclear what its actual motivation for changing the reward is. (What does "The human likely wants to see progress" mean? It doesn't exac...
2024-06-28T06:41:48.059Z
-6
wjZtzZfYi7Qn77H7k
FSgGBjDiaCdWxNBhj
Sycophancy to subterfuge: Investigating reward tampering in large language models
sycophancy-to-subterfuge-investigating-reward-tampering-in
https://arxiv.org/abs/2406.10162
Carson Denison
2024-06-17T18:41:31.090Z
iEu9kToRjMHN9hZDk
Yep, not claiming you did anything problematic, I just thought this selection might not be immediately obvious to readers and the random examples might be informative.
2024-06-28T15:27:20.004Z
4
5XpMLH7C4vpHpE3Ao
FSgGBjDiaCdWxNBhj
Sycophancy to subterfuge: Investigating reward tampering in large language models
sycophancy-to-subterfuge-investigating-reward-tampering-in
https://arxiv.org/abs/2406.10162
Carson Denison
2024-06-17T18:41:31.090Z
LvioqwQpbQcQuQfma
Hmm, I think I was wrong about DAgger and confused it with a somewhat different approach in my head. I agree that it provides bounds. (Under various assumptions about the learning algorithm that we can't prove for NNs but seem reasonable to assume in practice.) I now agree that the proposed method is basically just a...
2024-06-28T20:19:51.495Z
5
KCq8yfvrjEP2LhGcR
mXYdYh6L9odTJZDSm
An issue with training schemers with supervised fine-tuning
an-issue-with-training-schemers-with-supervised-fine-tuning
https://www.lesswrong.com/posts/mXYdYh6L9odTJZDSm/an-issue-with-training-schemers-with-supervised-fine-tuning
Fabien Roger
2024-06-27T15:37:56.020Z
vnxG9dHmt6S7tervb
> this failure mode is dangerous because of scheming AI and I say it's dangerous because the policy is OOD I would say that it is dangerous in the case where is is both OOD enough that the AI can discriminate and the AI is scheming. Neither alone would present a serious (i.e. catastrophic) risk in the imitation conte...
2024-06-28T20:21:51.202Z
2
KCq8yfvrjEP2LhGcR
mXYdYh6L9odTJZDSm
An issue with training schemers with supervised fine-tuning
an-issue-with-training-schemers-with-supervised-fine-tuning
https://www.lesswrong.com/posts/mXYdYh6L9odTJZDSm/an-issue-with-training-schemers-with-supervised-fine-tuning
Fabien Roger
2024-06-27T15:37:56.020Z
MsTX9LgtT5SrPPfm7
Thanks, I improved the wording.
2024-06-28T22:37:53.641Z
2
Skxp7ZdEQ44C6mHpz
mXYdYh6L9odTJZDSm
An issue with training schemers with supervised fine-tuning
an-issue-with-training-schemers-with-supervised-fine-tuning
https://www.lesswrong.com/posts/mXYdYh6L9odTJZDSm/an-issue-with-training-schemers-with-supervised-fine-tuning
Fabien Roger
2024-06-27T15:37:56.020Z
Dmio4qZtL3eWiEfu8
(I'll edit the post at some point to highlight this discussion and clarify this.)
2024-06-28T22:38:39.695Z
5
LvioqwQpbQcQuQfma
mXYdYh6L9odTJZDSm
An issue with training schemers with supervised fine-tuning
an-issue-with-training-schemers-with-supervised-fine-tuning
https://www.lesswrong.com/posts/mXYdYh6L9odTJZDSm/an-issue-with-training-schemers-with-supervised-fine-tuning
Fabien Roger
2024-06-27T15:37:56.020Z
gmPmwGsShenPWgwqP
I agree that a few factors of 2 don't matter much at all, but I think highlighting a specific low threshold relative to the paper seems misguided as opposed to generally updating based on the level of egregiousness and rarity. (Where you should probably think about the rarity in log space.) (I think I made the point t...
2024-06-28T23:32:29.830Z
6
yyiRY2eJQnn5GEMRh
FSgGBjDiaCdWxNBhj
Sycophancy to subterfuge: Investigating reward tampering in large language models
sycophancy-to-subterfuge-investigating-reward-tampering-in
https://arxiv.org/abs/2406.10162
Carson Denison
2024-06-17T18:41:31.090Z
bHzg5xnCJE5BwBwhG
Yep, reasonable summary. > I do think that it's important to keep in mind that for threat models where reward hacking is reinforced, small RTFs might matter a lot, but maybe that was already obvious to everyone. I don't think this was obvious to everyone and I appreciate this point - I edited my earlier comment to mo...
2024-06-29T00:17:49.225Z
4
yH6ePwtDquBpsTAQq
FSgGBjDiaCdWxNBhj
Sycophancy to subterfuge: Investigating reward tampering in large language models
sycophancy-to-subterfuge-investigating-reward-tampering-in
https://arxiv.org/abs/2406.10162
Carson Denison
2024-06-17T18:41:31.090Z
D3onCbwATt9LctMT9
(Also, to be clear, thanks for the comment. I strong upvoted it.)
2024-06-29T00:20:26.073Z
6
KCq8yfvrjEP2LhGcR
mXYdYh6L9odTJZDSm
An issue with training schemers with supervised fine-tuning
an-issue-with-training-schemers-with-supervised-fine-tuning
https://www.lesswrong.com/posts/mXYdYh6L9odTJZDSm/an-issue-with-training-schemers-with-supervised-fine-tuning
Fabien Roger
2024-06-27T15:37:56.020Z
iPG7okYhBLj6hXvys
Some more: - The AI kills a huge number of people with a bioweapon to destablize the world and relatively advantage its position. - Massive world war/nuclear war. This could kill 100s of millions easily. 1 billion is probably a bit on the higher end of what you'd expect. - The AI has control of some nations, but think...
2024-06-30T18:05:55.199Z
4
sFdxonpCtL2R5QfyE
tKk37BFkMzchtZThx
MIRI 2024 Communications Strategy
miri-2024-communications-strategy
https://www.lesswrong.com/posts/tKk37BFkMzchtZThx/miri-2024-communications-strategy
Gretta Duleba
2024-05-29T19:33:39.169Z
jxeej9rBdhMr2rxkJ
> Technically "substantial chance of at least 1 billion people dying" can imply the middle option there, but it sounds like you mean the central example to be closer to a billion than 7.9 billion or whatever. That feels like a narrow target and I don't really know what you have in mind. I think "crazy large scale con...
2024-06-30T18:07:35.978Z
2
bWSZDtLAhTmkCmWbh
tKk37BFkMzchtZThx
MIRI 2024 Communications Strategy
miri-2024-communications-strategy
https://www.lesswrong.com/posts/tKk37BFkMzchtZThx/miri-2024-communications-strategy
Gretta Duleba
2024-05-29T19:33:39.169Z
odfsZHDnPreyTYhba
Hmm, not exactly. Our verification ability only needs to be sufficiently good relative to the AIs.
2024-06-30T19:38:11.789Z
2
bJkoMeKZbP9ksfH5C
7fJRPB6CF6uPKMLWi
My AI Model Delta Compared To Christiano
my-ai-model-delta-compared-to-christiano
https://www.lesswrong.com/posts/7fJRPB6CF6uPKMLWi/my-ai-model-delta-compared-to-christiano
johnswentworth
2024-06-12T18:19:44.768Z
PrrwpfTYpkxNRqyzZ
(If the small black hole thing works out - it is non-obvious that this will be achievable even for technologically mature civilizations.)
2024-07-04T15:34:33.574Z
4
WhXeAfumfSxQHgkSL
EKhNDspRxLGcZAuJz
What percent of the sun would a Dyson Sphere cover?
what-percent-of-the-sun-would-a-dyson-sphere-cover
https://www.lesswrong.com/posts/EKhNDspRxLGcZAuJz/what-percent-of-the-sun-would-a-dyson-sphere-cover
Raemon
2024-07-03T17:27:50.826Z
LfmaNEKzHcPuutDqm
I don't think you can feasibly use the Hawking radiation of large black holes as an energy source in our universe (even if you are patient). My understanding is that larger black holes decay over ~$10^{100}$ years. I did a botec a while ago and found that you maybe get 1 flop every $10^{30}$ years or so on average if ...
2024-07-05T17:54:06.008Z
4
RTj257vAYhyHuvrc5
EKhNDspRxLGcZAuJz
What percent of the sun would a Dyson Sphere cover?
what-percent-of-the-sun-would-a-dyson-sphere-cover
https://www.lesswrong.com/posts/EKhNDspRxLGcZAuJz/what-percent-of-the-sun-would-a-dyson-sphere-cover
Raemon
2024-07-03T17:27:50.826Z
THWAwauecRthcKzQh
I think the size might have to be pretty precise to get this right (I think decay duration is cubic in mass), so they'd probably need to be engineered to have a particular size. (E.g. add mass to a small black hole until it hits the right size.) But, yeah, with this constraint, I think it can maybe work. (I don't know...
2024-07-06T03:15:15.317Z
2
cfzEWTKmHdMCyrTGz
EKhNDspRxLGcZAuJz
What percent of the sun would a Dyson Sphere cover?
what-percent-of-the-sun-would-a-dyson-sphere-cover
https://www.lesswrong.com/posts/EKhNDspRxLGcZAuJz/what-percent-of-the-sun-would-a-dyson-sphere-cover
Raemon
2024-07-03T17:27:50.826Z
sADrFvb8axdHTkZx7
Sure. Imagine you have an AI which was instructed to add some feature or fix some bug in a code base. That AI writes a bunch of code, runs various tests in a sandbox, and the submits a PR. We need to review this PR, rate how good it is, and determine if it is actively harmful (e.g. adds a vulnerability). The human c...
2024-07-06T19:58:24.094Z
10
YcgdckG5RLLhBKzhx
6AT4vhYzww56CR6cm
Scalable oversight as a quantitative rather than qualitative problem
scalable-oversight-as-a-quantitative-rather-than-qualitative
https://www.lesswrong.com/posts/6AT4vhYzww56CR6cm/scalable-oversight-as-a-quantitative-rather-than-qualitative
Buck
2024-07-06T17:42:41.325Z
BpX9vnaCk5TgZZqx7
I think a notion of understanding individual actions requires breaking things down into steps which aim to accomplish specific things. (Including potentially judging decompositions that AIs come up with.) So, in the maze case, you're probably fine just judging where it ends up (and the speed/cost of the path) given t...
2024-07-06T20:01:48.229Z
6
YcgdckG5RLLhBKzhx
6AT4vhYzww56CR6cm
Scalable oversight as a quantitative rather than qualitative problem
scalable-oversight-as-a-quantitative-rather-than-qualitative
https://www.lesswrong.com/posts/6AT4vhYzww56CR6cm/scalable-oversight-as-a-quantitative-rather-than-qualitative
Buck
2024-07-06T17:42:41.325Z
88eSpquacPTbR7KQ3
[Low confidence and low familiarity] My main issue with the case for singular learning theory is that I can't think of any particular use cases that seem both plausible and considerably useful. (And the stories I've heard don't seem compelling to me.) I think it seems heuristically good to generally understand more s...
2024-07-08T21:46:13.849Z
34
null
CmcarN6fGgTGwGuFp
Dialogue introduction to Singular Learning Theory
dialogue-introduction-to-singular-learning-theory
https://www.lesswrong.com/posts/CmcarN6fGgTGwGuFp/dialogue-introduction-to-singular-learning-theory
Olli Järviniemi
2024-07-08T16:58:10.108Z
NLLrmDe6yJv8JN9TL
Seems like a reasonable idea. To implement this, I'd have to look more carefully at exactly what types of mistakes GPT-4o makes to calibrate what should/shouldn't be dithered. (Additional programs are cheap, but you can easily get a combinatorial explosion with this sort of thing.) (I'm not currently working on ARC-AG...
2024-07-09T20:55:52.484Z
8
null
jEEWe42fcJWdbCZo9
Fix simple mistakes in ARC-AGI, etc.
fix-simple-mistakes-in-arc-agi-etc
https://www.lesswrong.com/posts/jEEWe42fcJWdbCZo9/fix-simple-mistakes-in-arc-agi-etc
Oleg Trott
2024-07-09T17:46:50.364Z
YGFjam9Nfaz8F4Bwy
> discrete phases, and the Developmental Landscape paper validates this Hmm, the phases seem only roughly discrete, and I think a perspective like the [multi-component learning](https://www.lesswrong.com/posts/RKDQCB6smLWgs2Mhr/multi-component-learning-and-s-curves) perspective totally explains these results, makes st...
2024-07-09T23:24:52.653Z
8
rFzw3pA4vyb9d9EEW
CmcarN6fGgTGwGuFp
Dialogue introduction to Singular Learning Theory
dialogue-introduction-to-singular-learning-theory
https://www.lesswrong.com/posts/CmcarN6fGgTGwGuFp/dialogue-introduction-to-singular-learning-theory
Olli Järviniemi
2024-07-08T16:58:10.108Z
SYLh8Nbt4jhzmfGgo
It sounds like your case for SLT that you make here is basically "it seems heuristically good to generally understand more stuff about how SGD works". This seems like a reasonable case, though considerably weaker than many other more direct theories of change IMO. > I think you might buy the high level argument for t...
2024-07-10T04:00:53.894Z
15
QotuTCqJawByDNr9T
CmcarN6fGgTGwGuFp
Dialogue introduction to Singular Learning Theory
dialogue-introduction-to-singular-learning-theory
https://www.lesswrong.com/posts/CmcarN6fGgTGwGuFp/dialogue-introduction-to-singular-learning-theory
Olli Järviniemi
2024-07-08T16:58:10.108Z
EGzzXuLGJBhx5RmhR
> At a high level the case for Bayesian statistics in alignment is that if you want to control engineering systems that are learned rather than designed, and if that learning means choosing parameters that have high probability with respect to some choice of dataset and model, then it makes sense to understand what the...
2024-07-10T04:20:16.477Z
12
QotuTCqJawByDNr9T
CmcarN6fGgTGwGuFp
Dialogue introduction to Singular Learning Theory
dialogue-introduction-to-singular-learning-theory
https://www.lesswrong.com/posts/CmcarN6fGgTGwGuFp/dialogue-introduction-to-singular-learning-theory
Olli Järviniemi
2024-07-08T16:58:10.108Z
uo8C2hX6Zxk8djfBo
Agree overall, but you might be able to use a notably cheaper model (e.g. GPT-3.5) to dither.
2024-07-10T21:54:39.103Z
2
39vv4CcSNj9zJLf8g
jEEWe42fcJWdbCZo9
Fix simple mistakes in ARC-AGI, etc.
fix-simple-mistakes-in-arc-agi-etc
https://www.lesswrong.com/posts/jEEWe42fcJWdbCZo9/fix-simple-mistakes-in-arc-agi-etc
Oleg Trott
2024-07-09T17:46:50.364Z
vHBDgLkdi3XESqvRx
No, but it doesn't need to spot errors, just note places which could plausibly be bugs.
2024-07-11T02:51:58.907Z
4
QGb7RFMxGBjbAz947
jEEWe42fcJWdbCZo9
Fix simple mistakes in ARC-AGI, etc.
fix-simple-mistakes-in-arc-agi-etc
https://www.lesswrong.com/posts/jEEWe42fcJWdbCZo9/fix-simple-mistakes-in-arc-agi-etc
Oleg Trott
2024-07-09T17:46:50.364Z
vRDxYE7e2v8FukpZP
Note that 90% of people struck by lightning survive, so that actual number struck per year is more like 300.
2024-07-16T01:17:46.910Z
8
nWaE4TZhvhh2PKG8B
R9nM8GuDWBubRKA7v
Will the growing deer prion epidemic spread to humans? Why not?
will-the-growing-deer-prion-epidemic-spread-to-humans-why
https://eukaryotewritesblog.com/2023/06/24/chronic-wasting-disease/
eukaryote
2023-06-25T04:31:56.824Z
iGoPYLBc8JgHGBzSD
Sadly, I don't think this proposal works. More precisely, testing if data is low perplexity is strictly worse than compressing the data against a small language model and then actually sending out a small payload. --- > Encoding this much data in a way that appears as natural language would be extremely challenging. ...
2024-07-22T22:31:45.662Z
4
null
aWZEDw6oxR6Wk5hru
Using an LLM perplexity filter to detect weight exfiltration
using-an-llm-perplexity-filter-to-detect-weight-exfiltration
https://www.lesswrong.com/posts/aWZEDw6oxR6Wk5hru/using-an-llm-perplexity-filter-to-detect-weight-exfiltration
Adam Karvonen
2024-07-21T18:18:05.612Z
kK4Gke4m2v3Fyfc6T
> Llama-3.1-405B not as good as GPT-4o or Claude Sonnet. Certainly Llama-3.1-70B is not as good as the similarly sized Claude Sonnet. If you are going to straight up use an API or chat interface, there seems to be little reason to use Llama. [Some providers are offering 405b at costs lower than 3.5 sonnet](https://art...
2024-07-24T19:59:49.447Z
10
null
fjzPg9ATbTJcnBZvg
Llama Llama-3-405B?
llama-llama-3-405b
https://www.lesswrong.com/posts/fjzPg9ATbTJcnBZvg/llama-llama-3-405b
Zvi
2024-07-24T19:40:07.565Z
swzE9e9xA5GwJFDmd
Huh? Why is it problematic to have your text edited by an LM? I do this all the time for minor clarity improvements and to fix typos. I agree that it would be rude to have a part or all of the comment be totally LM written without flagging this. In this case, seems plausible that the person asked gpt-4 to just summar...
2024-08-05T22:52:28.995Z
5
zceMwrMqGaLu7bMWg
fPvssZk3AoDzXwfwJ
Universal Basic Income and Poverty
universal-basic-income-and-poverty
https://www.lesswrong.com/posts/fPvssZk3AoDzXwfwJ/universal-basic-income-and-poverty
Eliezer Yudkowsky
2024-07-26T07:23:50.151Z
qfxBLnb3p3XruDnZd
> you should also be wary of 'minor clarity improvements' suggested by ChatGPT/Claude. I find a lot of them make prose worse, especially if you apply most of them so the gestalt becomes ChatGPTese. I agree with this, my editing process is to get the new output, view the diff and then copy in changes one-by-one if they...
2024-08-06T02:33:31.336Z
4
mbEtKcq3YvA3fzkjA
fPvssZk3AoDzXwfwJ
Universal Basic Income and Poverty
universal-basic-income-and-poverty
https://www.lesswrong.com/posts/fPvssZk3AoDzXwfwJ/universal-basic-income-and-poverty
Eliezer Yudkowsky
2024-07-26T07:23:50.151Z
uJ9jLtvA6GAqcCf5x
> A simple approach is to maintain intervals which are guaranteed to contain the actual values and to prove that output intervals don't overlap the unsafe region. For actual inference stacks in use (e.g. llama-3-405b 8 bit float) interval propagation will blow up massively and result in vacuous bounds. So, you'll mini...
2024-08-12T03:17:27.633Z
11
Bhp4e97NPBDfx3Anj
P8XcbnYi7ooB2KR2j
Provably Safe AI: Worldview and Projects
provably-safe-ai-worldview-and-projects
https://www.lesswrong.com/posts/P8XcbnYi7ooB2KR2j/provably-safe-ai-worldview-and-projects
Ben Goldhaber
2024-08-09T23:21:02.763Z
mKqk3qfJ6rNXvwenN
I agree you can do better than naive interval propagation by taking into account correlations. However, it will be tricky to get a much better bound while avoiding having this balloon in time complexity (all possible correlations requires exponentional time). More strongly, I think that if an adversary controlled the ...
2024-08-12T06:02:13.762Z
11
5qypian9DDFSaZEzY
P8XcbnYi7ooB2KR2j
Provably Safe AI: Worldview and Projects
provably-safe-ai-worldview-and-projects
https://www.lesswrong.com/posts/P8XcbnYi7ooB2KR2j/provably-safe-ai-worldview-and-projects
Ben Goldhaber
2024-08-09T23:21:02.763Z
T5czXsDjnsanKXRHk
> Rather, we require the AI to generate a program, control policy, or simple network for taking actions in the situation of interest. And we force it to generate a proof that it satisfies given safety requirements. If it can't do that, then it has no business taking actions in a dangerous setting. This seems near cer...
2024-08-12T06:07:56.473Z
12
5qypian9DDFSaZEzY
P8XcbnYi7ooB2KR2j
Provably Safe AI: Worldview and Projects
provably-safe-ai-worldview-and-projects
https://www.lesswrong.com/posts/P8XcbnYi7ooB2KR2j/provably-safe-ai-worldview-and-projects
Ben Goldhaber
2024-08-09T23:21:02.763Z
NeMQ5SNbycYvwDL5d
> Principle 1: Seek as broad and legitimate authority for your decisions as is possible under the circumstances I would have focused more on transparency and informing relevant groups rather than authority/accountability. (See also the comment from [Raemon](https://www.lesswrong.com/posts/aRciQsjgErCf5Y7D9/principles...
2024-09-01T23:14:35.545Z
16
null
aRciQsjgErCf5Y7D9
Principles for the AGI Race
principles-for-the-agi-race
https://www.lesswrong.com/posts/aRciQsjgErCf5Y7D9/principles-for-the-agi-race
William_S
2024-08-30T14:29:41.074Z
mRBQbrREf5zgd6Ewq
> Principle 2: Don’t take actions which impose significant risks to others without overwhelming evidence of net benefit > > [...] > > Significant margin of benefits over costs, accounting for possibility your calculations are incorrect (1.1x benefits over costs doesn’t justify, maybe 10x benefits over costs could justi...
2024-09-01T23:38:12.799Z
20
null
aRciQsjgErCf5Y7D9
Principles for the AGI Race
principles-for-the-agi-race
https://www.lesswrong.com/posts/aRciQsjgErCf5Y7D9/principles-for-the-agi-race
William_S
2024-08-30T14:29:41.074Z
rBa7yqRWFQbPDnoFi
> They could deeply understand the situations in which there's a treacherous turn, how the models decides whether to openly defect, and publish. It sounds as though you're imagining that we can proliferate the one case in which we caught the AI into many cases which can be well understood as independent (rather than ...
2024-09-02T19:23:53.489Z
6
T57EvmkcDmksAc4P4
YTZAmJKydD5hdRSeG
Would catching your AIs trying to escape convince AI developers to slow down or undeploy?
would-catching-your-ais-trying-to-escape-convince-ai
https://www.lesswrong.com/posts/YTZAmJKydD5hdRSeG/would-catching-your-ais-trying-to-escape-convince-ai
Buck
2024-08-26T16:46:18.872Z
xmLkoPrLdKnXbtQPu
Agreed, I really should have said "or possibly even train against it". I think SGD is likely to be much worse than best-of-N over a bunch of variations on the training scheme where the variations are intended to plausibly reduce the chance of scheming. Of course, if you are worried about scheming emerging thoughout tra...
2024-09-02T19:56:39.006Z
2
bLtJ9Ja426jaQZKjp
YTZAmJKydD5hdRSeG
Would catching your AIs trying to escape convince AI developers to slow down or undeploy?
would-catching-your-ais-trying-to-escape-convince-ai
https://www.lesswrong.com/posts/YTZAmJKydD5hdRSeG/would-catching-your-ais-trying-to-escape-convince-ai
Buck
2024-08-26T16:46:18.872Z
eCqWBjK4dgbpZ8pYT
Thanks for writing this! I appreciate the effort to make your perspective more transparent (and implicitly Anthropic's perspective as well). In this comment, I'll explain my two most important concerns with this proposal: - Fully state-proof security seems crucial at ASL-4, not only at ASL-5 - You should have a clear ...
2024-09-04T18:00:15.821Z
102
null
mGCcZnr4WjGjqzX5s
The Checklist: What Succeeding at AI Safety Will Involve
the-checklist-what-succeeding-at-ai-safety-will-involve
https://sleepinyourhat.github.io/checklist/
Sam Bowman
2024-09-03T18:18:34.230Z
7dE5aee5G5XFXdKHd
As far as security, perhaps part of what is going on is that you expect that achieving this high bar of security is too expensive: > ASL-4 is much more demanding and represents a rough upper limit on what we expect to be able to implement without heavily interfering with our research and deployment efforts. My sense ...
2024-09-04T18:35:16.119Z
2
eCqWBjK4dgbpZ8pYT
mGCcZnr4WjGjqzX5s
The Checklist: What Succeeding at AI Safety Will Involve
the-checklist-what-succeeding-at-ai-safety-will-involve
https://sleepinyourhat.github.io/checklist/
Sam Bowman
2024-09-03T18:18:34.230Z
SvRcdWDq37C9MpBD7
I left an earlier comment on more important and clearer points, but I thought I would also leave a comment with some more tentative remarks, explanations of where I'm confused, and notes. ## The story for ASL-4 safety seems unclear and/or unlikely to provide a "reasonable" level of safety In this checklist, perhaps t...
2024-09-04T19:33:36.641Z
49
null
mGCcZnr4WjGjqzX5s
The Checklist: What Succeeding at AI Safety Will Involve
the-checklist-what-succeeding-at-ai-safety-will-involve
https://sleepinyourhat.github.io/checklist/
Sam Bowman
2024-09-03T18:18:34.230Z
j7XyPeJwhJAwaQyWn
> My guess is that you think heavy government involvement should occur for before/during the creation of ASL-4 systems, since you're pretty concerned about risks from ASL-4 systems being developed in non-SL5 contexts. Yes, I think heavy government should occur once AIs can substantially accelerate general purpose R&D ...
2024-09-04T21:32:16.365Z
6
NvzqWxu8qxEPj4GZF
mGCcZnr4WjGjqzX5s
The Checklist: What Succeeding at AI Safety Will Involve
the-checklist-what-succeeding-at-ai-safety-will-involve
https://sleepinyourhat.github.io/checklist/
Sam Bowman
2024-09-03T18:18:34.230Z
nptLHHtGzkHopgtdK
Yeah, this all seems reasonable to me for the record, though I think any such proposal of this sort of norms needs to handle the fact that public discourse is sometimes very insane.
2024-09-04T21:33:39.852Z
9
dth4iaJL4hYpnJYCw
aRciQsjgErCf5Y7D9
Principles for the AGI Race
principles-for-the-agi-race
https://www.lesswrong.com/posts/aRciQsjgErCf5Y7D9/principles-for-the-agi-race
William_S
2024-08-30T14:29:41.074Z
fdKsWffbpdwDRFwac
It seems like I didn't do a good job of explaining the exit plan! I'll need to do a better job of explaining this in the future. (I'll respond to some of these specific points in a bit.)
2024-09-04T23:15:46.579Z
7
F9W5oxw536iAKnJwj
mGCcZnr4WjGjqzX5s
The Checklist: What Succeeding at AI Safety Will Involve
the-checklist-what-succeeding-at-ai-safety-will-involve
https://sleepinyourhat.github.io/checklist/
Sam Bowman
2024-09-03T18:18:34.230Z
eANGMjfpxNcrnBJvm
My proposal would roughly be that the US government (in collaboration with allies etc) enforces no one building AI which are qualitatively smarter than humans and this should be the default plan. (This might be doable without government support via coordination between multiple labs, but I basically doubt it.) Their ...
2024-09-05T05:52:27.614Z
7
52zzDYkT6bakC2mcN
mGCcZnr4WjGjqzX5s
The Checklist: What Succeeding at AI Safety Will Involve
the-checklist-what-succeeding-at-ai-safety-will-involve
https://sleepinyourhat.github.io/checklist/
Sam Bowman
2024-09-03T18:18:34.230Z
LpdiAW9eqb5wb4ixg
TBC, I don't think there are plausible alternatives to at least some US government involvement which don't require commiting a bunch of massive crimes. I have a policy against commiting or recommending commiting massive crimes.
2024-09-05T05:58:22.164Z
16
eANGMjfpxNcrnBJvm
mGCcZnr4WjGjqzX5s
The Checklist: What Succeeding at AI Safety Will Involve
the-checklist-what-succeeding-at-ai-safety-will-involve
https://sleepinyourhat.github.io/checklist/
Sam Bowman
2024-09-03T18:18:34.230Z
NLKdMRfDN33jfDiX6
> I don't know what set of beliefs implies that it's much more important to avoid building superhuman TAI once you have just-barely TAI, than to avoid building just-barely TAI in the first place. AIs which aren't qualitatively much smarter than humans seem plausible to use reasonably effectively while keeping risk dec...
2024-09-05T06:21:30.270Z
10
52zzDYkT6bakC2mcN
mGCcZnr4WjGjqzX5s
The Checklist: What Succeeding at AI Safety Will Involve
the-checklist-what-succeeding-at-ai-safety-will-involve
https://sleepinyourhat.github.io/checklist/
Sam Bowman
2024-09-03T18:18:34.230Z
gBKxjGhT6v7dSPZLs
Is your perspective something like: > With (properly motivated) qualitatively wildly superhuman AI, you can end the Acute Risk Period using means which aren't massive crimes despite not collaborating with the US government. This likely involves novel social technology. More minimally, if you did have a sufficiently al...
2024-09-05T17:15:22.263Z
4
pH3LkkfEHGzTfKMdP
mGCcZnr4WjGjqzX5s
The Checklist: What Succeeding at AI Safety Will Involve
the-checklist-what-succeeding-at-ai-safety-will-involve
https://sleepinyourhat.github.io/checklist/
Sam Bowman
2024-09-03T18:18:34.230Z
pr4iupfRuegzfKqbT
> if we get to just-barely TAI, at least not without plans that leverage that just-barely TAI in unsafe ways which violate the safety invariants of this plan I'm basically imagining being able to use controlled AIs which aren't qualitatively smarter than humans for whatever R&D purposes we want. (Though not applicatio...
2024-09-05T17:18:03.894Z
6
pH3LkkfEHGzTfKMdP
mGCcZnr4WjGjqzX5s
The Checklist: What Succeeding at AI Safety Will Involve
the-checklist-what-succeeding-at-ai-safety-will-involve
https://sleepinyourhat.github.io/checklist/
Sam Bowman
2024-09-03T18:18:34.230Z
HtPZda7KCBoKj8zeM
> How do you achieve such alignment? You wrote that you worry about the proposal of perfectly + scalably solving alignment, but I worry about how to achieve even the imperfect alignment of human-ish-level AIs that you're describing here. What techniques are you imagining using? I would say a mixture of moonshots and "...
2024-09-05T17:50:12.480Z
15
F9W5oxw536iAKnJwj
mGCcZnr4WjGjqzX5s
The Checklist: What Succeeding at AI Safety Will Involve
the-checklist-what-succeeding-at-ai-safety-will-involve
https://sleepinyourhat.github.io/checklist/
Sam Bowman
2024-09-03T18:18:34.230Z
eWmQzCgc39YmuMqTk
Presumably it helps a lot if AIs are clearly really powerful and their attempts to seize power are somewhat plausible? I think it's plausible (though not certain!) that people will take strong action *even if* they think the AI is role playing. I think this is a reasonable line to draw - people *should* freak out if ...
2024-09-05T18:26:01.552Z
13
L5e7CK5QASLigeev3
YTZAmJKydD5hdRSeG
Would catching your AIs trying to escape convince AI developers to slow down or undeploy?
would-catching-your-ais-trying-to-escape-convince-ai
https://www.lesswrong.com/posts/YTZAmJKydD5hdRSeG/would-catching-your-ais-trying-to-escape-convince-ai
Buck
2024-08-26T16:46:18.872Z
BcWcxR2trqoR23Rvy
> In particular, the model needs to bypass our control measures, which have been designed to secure against deliberate, best-effort attacks by the model, and the model needs to not get caught trying to do so during evaluations. Hopefully, but I wouldn't necessarily expect that we'll have reasonably well designed cont...
2024-09-06T18:21:41.500Z
12
EtLeQnM9fpPi9b9yG
YTZAmJKydD5hdRSeG
Would catching your AIs trying to escape convince AI developers to slow down or undeploy?
would-catching-your-ais-trying-to-escape-convince-ai
https://www.lesswrong.com/posts/YTZAmJKydD5hdRSeG/would-catching-your-ais-trying-to-escape-convince-ai
Buck
2024-08-26T16:46:18.872Z
q5vfGDSkZupyXWySQ
I raised a similar proposal to various people a while ago. The strongest objection I'm aware of is something like: > You actually want evaluators to have as much skin in the game as other employees so that when they take actions that might shut the company down or notably reduce the value of equity, this is a costly...
2024-09-07T22:57:44.516Z
15
null
sMBjsfNdezWFy6Dz5
Pay Risk Evaluators in Cash, Not Equity
pay-risk-evaluators-in-cash-not-equity
https://www.lesswrong.com/posts/sMBjsfNdezWFy6Dz5/pay-risk-evaluators-in-cash-not-equity
Adam Scholl
2024-09-07T02:37:59.659Z
SkuSNWx5pMkjppvhc
> Like, presumably the ideal scenario is that a risk evaluator estimates the risk in an objective way, and then the company takes (hopefully predefined) actions based on that estimate. The outcome of this interaction should not depend on social cues like how loyal they seem, or how personally costly it was for them to ...
2024-09-08T22:53:57.344Z
9
fm6q4p3uzFSNBwSHr
sMBjsfNdezWFy6Dz5
Pay Risk Evaluators in Cash, Not Equity
pay-risk-evaluators-in-cash-not-equity
https://www.lesswrong.com/posts/sMBjsfNdezWFy6Dz5/pay-risk-evaluators-in-cash-not-equity
Adam Scholl
2024-09-07T02:37:59.659Z
gJsyCLwszuZYkYjvj
I believe the approach is to forecast questions that resolve in the future and allow arbitrary internet access. I'm not totally sure though.
2024-09-10T22:00:49.582Z
7
ApExEZczHL4qALomD
4kuXNhPf9FBwok7tK
AI forecasting bots incoming
ai-forecasting-bots-incoming
https://www.safe.ai/blog/forecasting
Dan H
2024-09-09T19:14:31.050Z
23x5pbq9JHp7fZYtH
Daniel almost surely doesn't think growth will be constant. (Presumably he has a model similar to [the one here](https://www.lesswrong.com/posts/Gc9FGtdXhK9sCSEYu/what-a-compute-centric-framework-says-about-ai-takeoff).) I assume he also thinks that by the time energy production is >10x higher, the world has generally ...
2024-09-11T16:21:53.370Z
7
mkcqhxinj5oj2Pjou
K2D45BNxnZjdpSX2j
AI Timelines
ai-timelines
https://www.lesswrong.com/posts/K2D45BNxnZjdpSX2j/ai-timelines
habryka
2023-11-10T05:28:24.841Z